Home Artists Posts Import Register

Content

Get the TOX - www.patreon.com/posts/102788463

Important ! This version of StreamDiffusionTD replaces NDI + Spout with a faster solution for improved performance. However, I am planning to reintegrate ndi in the next version in a way that makes it easier to run non locally. 0.1.8 will only work locally on a single machine, but will be expanding support for non local.

[ EDIT - 0.1.9 Version fixes possible error when switching resolutions + issues with finding shared memory ]

Changelog:

  • Less latency, higher fps, and improved diffusion quality by replacing ndi+spout with direct memory buffer.

    • All handled inside of operator for you. If there are errors, perhaps change the 'Stream Out Name' in Stream Settings 2 and restart the stream.

  • You might notice a different curve on the denoising step sliders / T Index LIst.

    • It should be a bit more even through the steps and will better resemble input mage at higher values.

    • Improved input image coherence after diffusion

  • Added logic to automatically handle the memory i/o

    • Operator should handle loading output from StreamDiffusion better. No more disable/reenable needed.

In the works:

Many things I am testing / building with StreamDiffusion backend ... including tensorrt sdxl-turbo, prompt weighting, ip adapter images, controlnet, and also experimenting with stable diffusion + diffusers directly inside of TD. If you have any experience with any of these things and want to jam on some diffusers stuff, let me know !!

Enjoy ! 🧪

Comments

Ethan Alloway

Thanks! Excited to try this version.

Michele Di Pasquale

i have errors if i change the resolution to something different of 512x512 :/

Mikhail Kozlov

Any resolution different to 512x512 is not working. Acceleration is disabled.

Wingto

"Error in extracting image data: cannot reshape array of size 790528 into shape (800,800,3)"

melt.immersive

I can confirm - no resolution other than 512x512 works :/

Lyell Hintz

Just posted a fix for the resolution ! Found what was causing the issue definitely and tested across many resolutions. Note : will change resolution to multiplier of 8 pixels due to pipeline requirements.

melt.immersive

Hey! Any plans to incorporate Hyper-SD?

Lyell Hintz

It is working currently. I've been using the 4 step lora as a Custom LCM lora. I've done a bit of work on the model loading process as well. will be posting update soon. also dm me on discord and can share latest